DPA-4.0.1-Pro-MPtrj
Model overview
Interatomic potential: Energy with gradient-based forces and stress.
CPS 0.84 combines discovery, geometry optimization, and phonons. Inspect individual task results below.
Licenses are as declared by the submitter. Training dataset terms are listed separately below.
Model Info
Specifications
- Architecture
- gnn
- Outputs
- Energy with gradient-based forces and stress EFSG
- Openness
- Open source, open data OSOD
Discovery protocol
- Training
- structure to energy, force, stress S2EFS
- Evaluation
- initial structure to relaxed energy with structure relaxation IS2RE-SR
Training Set
MPtrj: 1.58M structures from 146k materials · MIT
Run this model
Use the submitted environment with the upstream instructions.
From a repository checkout, run uv run models/run_kappa.py --list-models to check checkpoint requirements. If this model requires one, add --checkpoint /path/to/checkpoint below. Then execute the printed command
for a phonon smoke test in the model's isolated environment.
uv run models/run_kappa.py --model dpa-4.0.1-pro-mptrj --print-cmd --dry-runSelect an available task above to inspect this model's results.
Model Authors
Trained By
description
DPA-4.0.1-Pro-MPtrj is the DPA4-Pro universal interatomic potential trained only on the MPtrj dataset for this Matbench Discovery submission.
architecture
DPA4 is an SE(3)-equivariant potential built on an EMFA (Edge-conditioned, Multi-Focus, Attention) SO(2)-equivariant convolution: a low-rank edge-node SO(2)-equivariant product, a multi-focus design for message nonlinearity, and envelope-gated attention for message aggregation, with a Lebedev-grid projection that preserves SO(3)-equivariance in the nonlinearity.
training
The model was trained for 2,000,000 steps on 16 GPUs using the HybridMuon optimizer, WSD learning-rate schedule, MAE energy/force/virial loss weights 20/20/5, with bf16 AMP, TF32 matmul, and torch compile enabled.